8

Problem 1: Let $x_i \ge 0, \, i=1, 2, \cdots, n$ with $\sum_{i=1}^n x_i = \frac12$. Prove or disprove that $$\sum_{1\le i < j \le n} \frac{x_ix_j}{1-x_i-x_j} \le \frac18.$$

This is related to the following problem:

Problem 2: Let $x_i \ge 0, \, i=1, 2, \cdots, n$ with $\sum_{i=1}^n x_i = \frac12$. Prove that $$\sum_{1\le i<j\le n}\frac{x_ix_j}{(1-x_i)(1-x_j)}\le \frac{n(n-1)}{2(2n-1)^2}.$$

Problem 2 is in "Problems From the Book", 2008, Ch. 2, which was proposed by Vasile Cartoaje.
See: Prove that $\sum_{1\le i<j\le n}\frac{x_ix_j}{(1-x_i)(1-x_j)} \le \frac{n(n-1)}{2(2n-1)^2}$

Background:

I proposed Problem 1 when I tried to find my 2nd proof for Problem 2.

It is not difficult to prove that $$\frac{1}{(2n-1)^4} + \frac{16n^2(n-1)^2}{(2n-1)^4}\cdot \frac{x_ix_j}{1-x_i-x_j} \ge \frac{x_ix_j}{(1-x_i)(1-x_j)}.$$ (Hint: Use $\frac{x_ix_j}{(1-x_i)(1-x_j)}= 1 - \frac{1}{1 + x_ix_j/(1-x_i-x_j)}$ and $\frac{1}{1+u} \ge \frac{1}{1+v} - \frac{1}{(1+v)^2}(u-v)$ for $u = x_ix_j/(1-x_i-x_j)$ and $v=\frac{1}{4n(n-1)}$. Or simply $\mathrm{LHS} - \mathrm{RHS} = \frac{(4x_ix_jn^2 - 4x_ix_j n + x_i + x_j - 1)^2}{(2n-1)^4(1-x_i-x_j)(1-x_i)(1-x_j)}\ge 0$.)

To prove Problem 2, it suffices to prove that $$\frac{1}{(2n-1)^4}\cdot \frac{n(n-1)}{2} + \frac{16n^2(n-1)^2}{(2n-1)^4}\sum_{1\le i < j \le n} \frac{x_ix_j}{1-x_i-x_j} \le \frac{n(n-1)}{2(2n-1)^2} $$ or $$\sum_{1\le i < j \le n} \frac{x_ix_j}{1-x_i-x_j} \le \frac18.$$

For $n=2, 3, 4$, the inequality is true.

For $n=5, 6$, numerical evidence supports the statement.

Any comments and solutions are welcome and appreciated.

nonuser
  • 91,557
River Li
  • 49,125
  • In the case $n=3$ it seems we have the inequalities :

    $$h\left(x,y\right)+h\left(y,z\right)+h\left(z,x\right)\leq 0.25\left(f\left(2\sqrt{xy}\right)+f\left(2\sqrt{yz}\right)+f\left(2\sqrt{zx}\right)\right)+0.75\frac{1}{8}\leq 0.125$$

    Where :

    $$h\left(x,y\right)=\frac{yx}{1-x-y},f\left(x\right)=\frac{\frac{x^{2}}{4}}{1-x}$$

    And obviously $x+y+z=0.5$ ,$x,y,z>0$

    I think it works also with the general case with the constraint above in the question .

    – Barackouda Jun 03 '22 at 08:40
  • @ErikSatie How did you prove your inequalities here? – River Li Jun 03 '22 at 09:42
  • I haven't a proof yet just some ideas but nothing concluding .If I have it I shall post it . But a question :Does the arithmetic compensating method works here ? – Barackouda Jun 03 '22 at 09:45
  • for example we have for $x,y>0$ and $0.5\ge x+y$ : $$f\left(\frac{\left(x+y\right)}{3}+\frac{4}{3}\sqrt{xy}\right)-h\left(x,y\right)\geq 0$$ and we have the inequality : $$0.125\ge f\left(\frac{\left(x+y\right)}{3}+\frac{4}{3}\sqrt{xy}\right)+f\left(\frac{\left(z+y\right)}{3}+\frac{4}{3}\sqrt{zy}\right)+f\left(\frac{\left(x+z\right)}{3}+\frac{4}{3}\sqrt{xz}\right)$$ for $x,y\in[0,0.18]$ – Barackouda Jun 03 '22 at 10:34
  • @ErikSatie I think arithmetic compensation method works but needs some work. – River Li Jun 03 '22 at 10:51
  • OK I found something else in the case $n=4$ we have as constraint :$$0\ge h\left(x,y\right)-f\left(1.5\sqrt{xy}\right)$$ where $f\left(x\right)=\frac{x^{2}}{2}$ and $h$ as above we have on the other hand : $$0.125\ge f\left(1.5\sqrt{xy}\right)+f\left(1.5\sqrt{xz}\right)+f\left(1.5\sqrt{zy}\right)+f\left(1.5\sqrt{dy}\right)+f\left(1.5\sqrt{dz}\right)+f\left(1.5\sqrt{dx}\right)$$ where $x,y,z,d>0$ and $x+y+z+d=0.5$ and $d\leq 0.3875$ – Barackouda Jun 03 '22 at 11:39
  • @ErikSatie You need to consider if your inequalities are easy to deal with. If your inequalities are more difficult than the original one, it does not make sense. – River Li Jun 03 '22 at 11:48
  • Using some substitution I got $$f\left(x\right)=\frac{a}{x-a-ax}$$ and at the equality case where $x=1/a$ there is a link with the golden ratio . Nice isn't it ? – Barackouda Jun 04 '22 at 09:16
  • Let $x,y\geq 0$ such that $x+y\leq 0.5$ then we have :

    $$\left(x+y\right)^{2}\left(1-\left(x+y\right)\right)\geq \frac{xy}{1-x-y}$$

    then as the function $f(x)=x^2(1-x)$ is convex on $[0,1/3]$ we use Jensen-Mercer inequality to get :

    $$0.125\ge\frac{n\left(n-1\right)}{2}\left(f\left(x\right)+f\left(y\right)-f\left(x+y-\frac{1}{3}\right)\right)$$

    Wich is a constraint .

    Mercer, A.McD.. "A variant of Jensen's inequality.." JIPAM. Journal of Inequalities in Pure & Applied Mathematics [electronic only] 4.4 (2003): Paper No. 73, 2 p.-Paper No. 73, 2 p.

    – Barackouda Jun 04 '22 at 09:18

3 Answers3

6

Write $p_i = 2x_i$ and note that $\sum_i p_i = 1$. Then

\begin{align*} 1 + \sum_i \frac{p_i^2}{1 - p_i} &= \sum_i \frac{p_i}{1 - p_i} \\ &= \sum_{i,j} \frac{1}{2} \left( \frac{1}{1 - p_i} + \frac{1}{1 - p_j} \right) p_i p_j \\ &\geq \sum_{i,j} \left( \frac{2}{2-p_i-p_j} \right) p_i p_j. \tag{by AM–HM} \end{align*}

Rearranging this inequality, we get

$$ 1 \geq \sum_{i \neq j} \frac{2p_i p_j}{2 - p_i - p_j} = 8 \sum_{i < j} \frac{x_i x_j}{1 - x_i - x_j},$$

completing the proof.

Sangchul Lee
  • 181,930
  • 1
    Thanks! It is impressive. – River Li Jun 04 '22 at 22:50
  • @RiverLi, Glad it helped! :) – Sangchul Lee Jun 04 '22 at 23:09
  • The key step is the AM-HM step which is simple but nice. It is a common trick $\frac{4}{x + y} \le 1/x + 1/y$ (C-S form). – River Li Jun 04 '22 at 23:21
  • Sorry, I don't see the second equality. For example, if $p_1=\frac13$ and $p_2=\frac23$, then $$\sum_i \frac{p_i}{1 - p_i}=\frac52$$ and $$ \sum_{i,j} \frac{1}{2} \left( \frac{1}{1 - p_i} + \frac{1}{1 - p_j} \right) p_i p_j =\frac12.$$ – John Bentin Jun 18 '22 at 16:44
  • @John, I guess you only computed the term corresponding to $(i,j)=(1,2)$ in the last sum you quoted. – Sangchul Lee Jun 18 '22 at 21:47
  • I chose $n=2$ and computed the sum for $1\leqslant i<j\leqslant n$, as specified in the question. – John Bentin Jun 19 '22 at 11:14
  • @JohnBentin, Your sum is missing the diagonal terms. Note that the sum in your case reduces to \begin{align}&\frac{1}{2}\left(\frac{1}{1-p_1}+\frac{1}{1-p_1}\right)p_1p_1+\frac{1}{2}\left(\frac{1}{1-p_1}+\frac{1}{1-p_2}\right)p_1p_2\&\quad+\frac{1}{2}\left(\frac{1}{1-p_2}+\frac{1}{1-p_1}\right)p_2p_1+\frac{1}{2}\left(\frac{1}{1-p_2}+\frac{1}{1-p_2}\right)p_2p_2,\end{align} which then evaluates to $\frac{5}{2}$ as expected. – Sangchul Lee Jun 19 '22 at 14:14
  • Thank you, Sangchul. I did not take in that you were switching from the summation domain ${(i,j): 1\leqslant i<j\leqslant n}$ to the domain ${1,...,n}^2$. I should have read your notation $\sum_{i,j}$ according to its precise meaning. – John Bentin Jun 19 '22 at 14:49
  • @JohnBentin, No worries! :) – Sangchul Lee Jun 19 '22 at 15:05
0

Partial proof with hint :

Let us consider the inequality for $x,y>0$ and $x,y\leq 0.5$ :

$$l\left(x,y\right)=g\left(2xy+\frac{4}{3}xy\left(x-y\right)^{2}\right)-\frac{x^2y^2}{1-x^2-y^2}\ge 0$$

Where $g\left(x\right)=\frac{\frac{x^{2}}{4}}{1-x}$

As $g$ is convex on $[0,0.5]$ we use Jensen-Mercer inequality

We need to show :

$$\frac{n\left(n-1\right)}{2}\left(h\left(x,y\right)+h\left(a,b\right)-g\left(f\left(x,y\right)+f\left(a,b\right)-d\right)\right)\le\frac{1}{8}\tag{I}$$

Where $f\left(x,y\right)=2xy+\frac{4}{3}xy\left(x-y\right)^{2},h\left(x,y\right)=g\left(f\left(x,y\right)\right)$ and $f(a,b)\leq d \leq f\left(x,y\right)$

Now we use the inverse function of $f(x)$ with a positive value or :

$$r\left(x\right)=2\sqrt{x^{2}+x}-2x$$

So $(I)$ is true with the constraint (where we have used Karamata's inequality)

$$f\left(x,y\right)+f\left(a,b\right)\le r\left(\frac{\frac{1}{8}2}{n\left(n-1\right)}+g\left(f\left(x,y\right)+f\left(a,b\right)-d\right)\right)+r(0)$$

Barackouda
  • 3,879
  • $\frac{xy}{1-x-y}-f\left(\frac{4}{\frac{1}{x}+\frac{1}{y}}\right)\le0$ seems not true for $x = 1/500, y = 1/100$. – River Li Jun 04 '22 at 12:05
  • @RiverLi Read carefully I mean here "constraint". – Barackouda Jun 04 '22 at 13:39
  • I don't understand your writing. – River Li Jun 04 '22 at 13:48
  • @RiverLi I rewrite it and add some details see please . – Barackouda Jun 04 '22 at 15:55
  • @RiverLi Hi I do the most difficult what do you think about ? There are the two equality case . Hope you enjoy it . – Barackouda Jun 18 '22 at 17:48
  • Sorry, It is hard for me to understand your answer. For example, "where we have used the condition as", "Then we use the inverse function of f(x) with a positive value or $r(x)=\cdots$", "And where we assume that there some unknows equals to zero". – River Li Jun 18 '22 at 22:20
  • I don't follow for example, "We need to show ... $f(a,b)\leq d \leq f\left(x,y\right)$". Why does it suffice to prove (I)? You should give details. – River Li Jun 19 '22 at 23:22
  • You should use punctuation marks such as comma, period. For example, "where $g\left(x\right)=\frac{\frac{x^{2}}{4}}{1-x}$(period or comma here)". – River Li Jun 19 '22 at 23:39
  • By the way, "Where ..." is not a new sentence, why did you capitalize Where? Also, do you mean "As $g$ is convex on $[0,0.5]$, ... we need to show .. "? If so, why did you capitalize We? – River Li Jun 19 '22 at 23:44
0

An incomplete approach:

$\sum\limits_{1\le i < j \le n} A_i A_j=\frac{1}{2}[(\sum_{k=1}^{n} A_k )^2=\sum_{k=1}^{n} A_k^2]=$

Using integral representation, we can write $\sum\limits_{1\le i < j \le n} \frac{x_ix_j}{1-x_i-x_j} =\sum\limits_{1\le i < j \le n} \int_{0}^{1} x_i x_j t^{-x_i-x_j} dt=\int_{0}^{1} dt\sum\limits_{1\le i < j \le n} x_i t^{-x_i} ~x_j t^{-x_j}= \frac{1}{2} [(\sum_{k=1}^{n} \int_{0}^{1} x_k t^{-x_k})^2+\sum_{k=1}^{n} \int_{0}^{1} x_k^2 t^{-2x_k}]dt= \frac{1}{2} \left[\frac{x_k^2}{(1-x_k)^2}-\frac{x_k^2}{1-2x_k}\right]$

I wish to come back again......

Z Ahmed
  • 46,319